fix: remind supervisors to refresh backlog after teardown

[kunchenguid]

Intent

Add a plain backlog-refresh reminder to firstmate so queued-item readiness stops going stale. The requested scope is intentionally minimal: after successful fm-teardown.sh completion, print a plain stdout reminder telling the supervisor to update data/backlog.md, move the finished id to Done, keep Done to the 10 most recent, re-scan Queued for items unblocked by the finished id or now time-due, and dispatch ready work. Also tighten one AGENTS.md backlog/supervision line to explicitly include time/date-gated queued items whose date has arrived. Do not add grep or auto-scan logic, structured ready-when or blocked-by machine fields, cron jobs, or new scripts.

What Changed

• Updates bin/fm-teardown.sh to print a post-teardown reminder to refresh data/backlog.md, prune Done entries, re-scan queued work, and dispatch anything ready.
• Tightens firstmate supervision docs in AGENTS.md to call out time/date-gated queued items whose date has arrived.
• Keeps the README teardown guidance aligned with the new backlog-refresh reminder.

Risk Assessment

✅ Low: Captain, the change is a narrow reminder/documentation update that runs only after the existing teardown path succeeds and does not alter teardown safety logic.

Testing

Captain, I inspected the intent, ran the existing shell behavior tests, then exercised the successful fm-teardown.sh path with reviewer-visible stdout evidence; stderr was empty, transient worktree fixtures were removed, and the working tree was clean afterward.

Evidence: fm-teardown reminder transcript

$ bin/fm-teardown.sh test-backlog-r3 teardown test-backlog-r3 complete ... 🌱 Backlog: test-backlog-r3 just finished. Update data/backlog.md - move test-backlog-r3 to Done (keep Done to the 10 most recent), then re-scan Queued for items now unblocked ... or now time-due, and dispatch what's ready.

$ bin/fm-teardown.sh test-backlog-r3
teardown test-backlog-r3 complete (window fm-test-backlog-r3, worktree /Users/kunchen/.no-mistakes/worktrees/016d88035d58/01KVNPEVJAYB1NBBC5A3KT98YG/.test-tmp/test-backlog-r3/worktree)
🌱 Backlog: test-backlog-r3 just finished. Update data/backlog.md - move test-backlog-r3 to Done (keep Done to the 10 most recent), then re-scan Queued for items now unblocked (a "blocked-by: test-backlog-r3" may have just cleared) or now time-due, and dispatch what's ready.
$ test ! -e state/test-backlog-r3.meta && echo "state cleanup: meta removed"
state cleanup: meta removed
$ cat .test-tmp/test-backlog-r3/treehouse-call.log
treehouse return --force /Users/kunchen/.no-mistakes/worktrees/016d88035d58/01KVNPEVJAYB1NBBC5A3KT98YG/.test-tmp/test-backlog-r3/worktree

Evidence: AGENTS backlog instruction snippet

AGENTS.md line 489: Re-evaluate Queued on every teardown and every heartbeat: anything whose blocker is gone gets dispatched, and time/date-gated items whose date has arrived get dispatched too.

   473	`data/backlog.md` is the durable queue.
   474	Update it on every dispatch, completion, and decision.
   475	
   476	`` `markdown
   477	## In flight
   478	- [ ] <id> - <one line> (repo: <name>, since <date>)
   479	
   480	## Queued
   481	- [ ] <id> - <one line> (repo: <name>) blocked-by: <id> - <reason>
   482	
   483	## Done
   484	- [x] <id> - <one line> - <https://github.com/owner/repo/pull/number> (merged <date>)
   485	- [x] <id> - <one line> - local main (merged <date>)
   486	- [x] <id> - <one line> - data/<id>/report.md (reported <date>)
   487	`` `
   488	
   489	Re-evaluate Queued on every teardown and every heartbeat: anything whose blocker is gone gets dispatched, and time/date-gated items whose date has arrived get dispatched too.
   490	
   491	Keep Done to the 10 most recent entries; prune older ones whenever you add to the section.

Pipeline

Updates from git push no-mistakes

✅ **intent** - passed

✅ No issues found.

✅ **Rebase** - passed

✅ No issues found.

✅ **Review** - passed

✅ No issues found.

✅ **Test** - passed

✅ No issues found.

• tests/fm-wake-queue.test.sh
• Created a temporary scout-style teardown fixture for test-backlog-r3, ran bin/fm-teardown.sh test-backlog-r3 with a fake treehouse return --force on PATH, captured stdout/stderr, and grepped stdout for the completion line plus the Done pruning, blocked-by, time-due, and dispatch reminder text.
• nl -ba AGENTS.md | sed -n '473,491p' > /var/folders/5x/4nqprlbx0518k3ybcb1sz6gr0000gn/T/no-mistakes-evidence/01KVNPEVJAYB1NBBC5A3KT98YG/agents-backlog-instruction.txt plus a targeted grep -F for the time/date-gated queued-item instruction.

✅ **Document** - passed

✅ No issues found.

✅ **Lint** - passed

✅ No issues found.

✅ **Push** - passed

✅ No issues found.

[bpsf]

Independent review of head a92a5362

Verdict: mergeable with one listed open item.

The PR delivers the story: explicit --mode direct-pr, no-mistakes unchanged as default, run optional only in direct-PR mode, run validation still refused in pipeline mode, byte-identical rendering on the existing pipeline path, honest labels, forge-head binding, sanitization, unknown timing, and an honest mixed-mode aggregate. No schema bump is needed because identity.no_mistakes_run already existed and main never read it.

Finding 1 (medium): direct-PR receipt records a ci_step duration the report says does not exist

• bin/fm-delivery-report.sh: the direct-PR override (ci_step=unknown; parked=unknown) lives only in render_report (~L521). write_receipt (~L735) recomputes ci_step=$(duration_or_status ci_step ci) with no mode branch.
• Trigger: render --mode direct-pr --timeline t.json where the timeline has no no_mistakes_run but its stages include {"name":"ci_step","duration_seconds":2538}. The script accepts it.
• Observed: report row | CI step | unknown - direct-PR has no no-mistakes pipeline stage |; receipt "ci_step_seconds":"2538" with "no_mistakes_run":"unknown". The receipt is what aggregate consumes, so a direct-PR receipt carries a pipeline-stage duration under a report that denies one exists.
• Coverage: the new direct-PR test uses no timeline, so deleting the override entirely leaves both new tests green (mutation probe).
• Smallest remedy: add the same mode branch in write_receipt (or hoist both lines into one helper used by both), and add one assertion to the direct-PR test with a timeline that carries a ci_step stage: report CI step | unknown and receipt .durations.ci_step_seconds == "unknown". If the timeline should win instead, apply that in both outputs and reconcile the report note. Either way the two outputs must agree.
• Today aggregate does not read ci_step_seconds, so no current aggregate line is wrong; the receipt is.

Verified (suite + targeted probes only, from disposable copies of main and this head)

• tests/fm-delivery-report.test.sh: 38 ok at this head; main also green.
• Byte identity vs main (cmp): minimal-args render (report + receipt), explicit --mode no-mistakes render, baseline-timeline render (report + receipt), publish against an attested body (published body + stdout). All identical.
• Direct-PR render/publish without --run: succeeds, fake no-mistakes never invoked, Validation: direct-PR, receipt schema still fm-delivery-receipt-v1.
• Forge-head binding in direct-PR mode: stale forge head refused, matching head publishes with Publication binding: GitHub forge head; a CI run on a different head is refused; hostile gate fixture leaves no secret or path in report or body.
• Pipeline mode refusals: no --run, timeline without no_mistakes_run, --run unknown, timeline no_mistakes_run: "unknown" all refused, so "unknown" in a receipt can only come from direct-PR mode.
• Aggregate over pipeline + direct-PR + legacy receipt lacking the field: no-mistakes=1 · direct-PR=1 · unknown=1; main's aggregate still reads the same receipts.
• Mutation probes against the two new tests: reverting the minimal-timeline run requirement, the collect guard, the identity gate, --mode parsing, the aggregate line, or the direct-PR label each fails a test. Removing the render-side ci_step/parked override does not (finding 1).
• docs/scripts.md row and the script header are accurate.

Full report with commands and outputs is recorded in the Firstmate home (data/fm-29-review/report.md).

Independent review (Firstmate scout, cursor-bugbot-review)